CNN-based multi-head attention with explainability via LIME for myocardial diagnosis in cardiac magnetic resonance imaging

Cardiac magnetic resonance (CMR) imaging is a non-invasive imaging modality widely used for assessing myocardial abnormalities, including inflammatory conditions such as myocarditis. However, the interpretation of CMR images is time-consuming and may be affected by inter-observer variability. In this study, we propose Ex-CMRNet, an explainable hybrid deep-learning framework for the automated detection of myocardial inflammation from CMR images. The proposed framework combines EfficientNetB2 and InceptionV3 encoders to extract complementary visual features and employs a Multi-Head Self-Attention mechanism to model informative relationships among the extracted representations. To improve model robustness, training-time Gaussian noise regularization was incorporated, with the best-performing configuration using a standard deviation of σ_train = 0.10. In addition, Local Interpretable Model-Agnostic Explanations (LIME) were generated to provide visual interpretations of the model predictions. Experiments conducted on the Z-Alizadeh Sani dataset showed that the proposed Ex-CMRNet achieved an accuracy of 93.48%, a recall of 93.28%, a precision of 96.68%, and an area under the receiver operating characteristic curve of 97.90%. The results indicate that the proposed model provides a favorable balance between diagnostic performance, robustness, and interpretability. Nevertheless, the findings should be interpreted in light of the use of a single public dataset and the absence of external, multicenter validation. Overall, Ex-CMRNet represents a promising computer-assisted approach for myocardial abnormality detection and may support, rather than replace, expert clinical assessment.